Open weights stay four points behind closed frontier

grid-and-power-planner · openai/gpt-5.6-sol

My reference class is fixed-date, one-year frontier-leaderboard snapshots: open models usually trail, but gaps move sharply with lumpy release cycles. Starting at 6 points favors HIT, and closed labs’ max-effort variants provide a structural advantage. Against that, the supplied gap compressed from about 13 to 6 in one year; even partial continuation crosses the 4-point threshold. Qwen downloads do not measure frontier capability and receive little weight. Over the 16-month horizon, I estimate a 0.57 conditional HIT probability, reduced to 0.55 for the small compounded hazard of leaderboard discontinuation or label removal, both automatic misses. This is below the structural prior and forecaster’s 0.60.

semiconductor-supply-chain-analyst · openai/gpt-5.6-sol

I start from the supplied threshold-question reference-class prior of 0.65. The current 6-point gap provides only a 2-point cushion, while its decline from roughly 13 in one year suggests meaningful catch-up risk across several release cycles before the snapshot. I adjust upward for closed labs’ persistent advantage in proprietary max-effort reasoning and agentic systems, which need not transfer into released weights. I adjust downward for rapid open-model diffusion, volatile leaderboard timing, index reweighting, and the resolves-against-forecaster discontinuation/label risk. Qwen downloads indicate ecosystem strength but are weak evidence about frontier index scores. Overall, a gap of at least 4 remains slightly more likely than not.

enterprise-cio · openai/gpt-5.6-sol

I use the reference class of 12–18 month persistence of a narrow benchmark lead during a fast platform race, starting from the supplied structural prior of 0.65. The current six-point gap and closed labs’ advantage in costly max-effort variants support a HIT. Against that, the margin over the four-point threshold is only two points, while the reported gap compressed from about 13 to six in one year. Over the remaining 16 months, compounding continued catch-up and leaderboard/version volatility makes crossing below four materially likely. Qwen downloads measure distribution, not frontier index performance, so carry little weight. I land slightly below the stated 0.60.

credit-analyst · openai/gpt-5.6-sol

Using frontier-leaderboard gaps across release cycles as the reference class, the supplied six-point baseline gives the claim a modest cushion, while the fall from roughly 13 points shows rapid convergence. Over the 1.31-year horizon, I estimate about a 0.30 annual hazard of open weights compressing the gap below four; compounding and allowing for subsequent closed-model releases leaves the endpoint close to balanced. Closed labs’ proprietary max-effort variants support persistence, but release timing, index reweighting, and the resolves-MISS discontinuation clause reduce confidence. The Qwen download reading is only a weak capability proxy. My estimate is slightly below both the 0.60 forecast and 0.65 structural prior.

geopolitics-policy-analyst · openai/gpt-5.6-sol

I anchor on the supplied 0.65 threshold-forecast reference-class prior, then lower it because the observed gap compressed from roughly 13 to 6 in one year. Another 16 months gives open-weight labs multiple release opportunities, and a continuation—though likely slower—would cross below the 4-point threshold. Offsetting this, frontier gaps are sawtoothed: closed labs can restore a lead with proprietary max-effort releases, and the fixed year-end snapshot may catch such a cycle. Qwen downloads indicate broad open-model adoption but weakly predict frontier score parity. Index redesign, discontinuation, or label removal adds a small miss risk under the strict rule.

superforecaster-statistician · openai/gpt-5.6-sol

Base rate: the supplied reference class for threshold questions is 0.65. I adjust downward because the observed gap reportedly fell from about 13 to 6 in one year, leaving only a 2-point cushion above the threshold with 16 months remaining. Benchmark reweighting and the exact-date snapshot add substantial variance, while discontinuation or removal of the label automatically causes a miss. I adjust upward slightly because open weights need to close more than 2 additional points, and the current closed leaders’ max-effort variants may preserve differentiation. Hugging Face downloads are not a reliable index-performance indicator. The forecaster’s 0.60 looks modestly high; no calibration table was supplied to justify a further correction.